Ant Group’s Robbyant Open-Sources LingBot-Vision: A 1B Boundary-Centric Vision Foundation Model for Dense Spatial Perception

A vision foundation model with roughly 1 billion parameters has been open-sourced for tasks requiring precise spatial understanding, such as identifying object boundaries and depth changes. Unlike most vision models that learn general image content while discarding fine-grained spatial details, this model prioritizes boundaries as a core training signal, allowing it to match or exceed the performance of models up to 7 times larger on dense spatial tasks. The approach uses masked boundary modeling during self-supervised training, where the model learns to recover masked image patches by treating boundary regions as especially informative and routing them through a dedicated geometric target alongside semantic signals. On benchmarks testing depth estimation and semantic segmentation, the model achieves competitive results with models substantially larger while consuming training data and computational resources significantly smaller than comparable approaches.